Telus Is Erasing Offshore Workers' Accents With AI. Three Questions Before You Deploy Something Similar.

If you have authority to roll out AI tools to a team — not just for your own work, but onto the people in your org — you face a different decision than deploying AI on data or documents. Telus's accent tool is a case study in what happens when that decision isn't made deliberately.

What Telus did

Telus Digital deployed real-time voice modification on offshore call-centre agents to alter their accents during live customer calls — without disclosing this to customers (Hacker News, May 6, 235 upvotes). Labour unions representing the affected workers called it deceptive. Rogers and Bell, two of Canada's other major telecoms, declined to implement the same tool when given the opportunity. The OECD flagged it as an AI ethics incident.

Telus's position is that the tool improves comprehension and reduces miscommunication. That may be true. But two competitors looked at the same technology and said no. The gap between "it works technically" and "we should deploy it" is where most AI ethics incidents start.

Three questions before you deploy

These aren't philosophical exercises. They're the questions Rogers and Bell probably worked through before declining.

1. Would the person whose behaviour the AI is modifying consent if you asked them directly? Telus's agents were not asked — they were enrolled. If the people affected would say yes when asked, asking costs nothing. If they would say no, that is information you need before rollout, not after.

2. Would the people on the receiving end change their behaviour if they knew? Customers on Telus calls didn't know accent modification was running. Some would feel deceived if they found out; some would not care. But "they probably wouldn't mind" is not the same as "we told them." If you would not want the deployment disclosed in a press release, that is a signal worth taking seriously before launch.

3. Are you deploying this reversibly? When Rogers and Bell said no, they preserved the option to revisit later. Unwinding an undisclosed deployment — after a union complaint, an OECD flag, or a press inquiry — is significantly harder than winding down a transparent one. Reversibility is not caution for its own sake. It is risk management.

The deployment that changed shape

Danielle runs customer operations at a 280-person logistics SaaS. Her team fielded a vendor pitch for an AI writing tool that would automatically rewrite support agents' replies into "brand-standard" English before sending them to customers. Agents would see their original draft; customers would receive the rewritten version. No disclosure in the reply thread.

She ran through the three questions. On the first: she was not confident her agents would consent if asked. Some would find it useful; others would feel their voice was being replaced. On the second: if customers knew they were reading AI-rewritten replies rather than words from the named agent on the thread, some would adjust their trust level. On the third: the tool was designed as a permanent background process, with no per-agent opt-out.

She didn't reject AI-assisted writing. She changed the rollout. Agents can opt in, they see the rewritten version before it sends, and the tool is disclosed in the support team's standard email footer. Same technology. Different deployment decision. The difference is that this version can be defended out loud.

When the three questions don't give a clean answer

Kevin is a people-ops director at a 65-person professional services firm evaluating AI-scored video interviews — a tool that watches candidate recordings and rates communication clarity, structured reasoning, and confidence markers.

The three questions don't produce a clean yes or no here. Candidates know they're being recorded but don't know the AI is scoring them. The vendor claims bias testing was conducted, but Kevin cannot verify the methodology. The tool would reduce review burden on three senior partners who currently watch every video.

His interim position: disclose the AI scoring in the candidate intake form before recording starts, add it to the firm's interview process FAQ, and run AI scores alongside — not instead of — partner reviews for the first two cycles. More cumbersome than flipping the tool on. But it is a deployment he can explain to a candidate who asks how they were assessed, and it gives the firm real data before committing.

The question Rogers and Bell answered

If you wouldn't be comfortable explaining the deployment to the people it affects — or to the customers on the other end — you haven't finished making the decision.